WHoW Framework for Moderation Analysis
Chen, Frermann & Lau, 2024 — University of Melbourne
How do you systematically describe what a moderator does? This paper from the University of Melbourne proposes a three-dimensional framework: Why (what’s the motive?), How (what dialogue act?), and Who (who’s being addressed?). The framework lets researchers compare moderation behavior across domains and could serve as a foundation for training or evaluating AI moderators.
The Three Dimensions
Why captures the underlying purpose of an intervention. Is the moderator trying to get information flowing (informational)? Keep things on track and on time (coordinative)? Manage emotions and interpersonal dynamics (social)? A single utterance can serve multiple motives—asking someone to respond to criticism both gets information and manages the social dynamic.
How describes the mechanism of intervention. Probing prompts someone to speak. Confronting asks one participant to respond to another’s point. Instruction directly shapes behavior. Interpretation reframes, summarizes, or connects to earlier discussion. Supplement adds information without pushing for behavioral change. These categories are mutually exclusive—each utterance gets one dialogue act label.
Who identifies the target: a specific individual, everyone in the room, the audience, or sometimes the moderator talking to themselves (thinking aloud).
What the Data Shows
The researchers annotated over 5,000 sentences from TV debates (Intelligence Squared) and NPR panel discussions. The domains show strikingly different moderation styles:
Debate moderators are primarily coordinative (66% of interventions). They use instructions frequently, facilitate direct interaction between participants through confronting questions, and work to manage time and flow.
Panel moderators are primarily informational (72% of interventions). They participate actively in the discussion, add supplementary information, and focus less on fostering participant-to-participant exchange.
This isn’t better or worse—it reflects different goals. Debates aim to produce clash between positions. Panels aim to surface expertise.
Measuring Moderation Style
The paper introduces three metrics that could inform evaluation frameworks:
Pro-activity — how often does the moderator initiate rather than respond? (measured by whether the target differs from the previous speaker)
Interactivity — do participants actually respond when addressed? (measured by whether the target becomes the next speaker)
Specificity — does the moderator address individuals or the whole group?
A highly proactive, highly specific moderator actively steers discussion. A reactive, group-addressed moderator lets things flow. Neither is inherently superior—it depends on what kind of conversation you’re running.
Connection to OFL
WHoW maps surprisingly well onto OFL’s Why-How-Who framework:
| WHoW | OFL |
|---|---|
| Why (Motives) | Why (Purpose, outcomes) |
| How (Dialogue Acts) | How (Intervention style, techniques) |
| Who (Target Speaker) | Who (Roles, participant dynamics) |
The main difference: WHoW operates at utterance level (annotating individual moderator statements), while OFL’s framework operates at methodology level (characterizing entire facilitation patterns). WHoW could provide the fine-grained behavioral data that validates or challenges OFL’s higher-level claims.
The Dataset
The researchers released both human annotations (5,657 sentences) and GPT-4o annotations (15,494 sentences) with a high inter-annotator agreement (Krippendorff’s alpha 0.37-0.75 depending on dimension). Code and data are available on GitHub.
Related Research
The Fora corpus provides complementary data on facilitation (a related but distinct practice). For the OFL framework inspired by this research, see Why-How-Who. The Korre et al. survey situates WHoW within broader discussion quality research.